U.S. patent application number 14/129702 was filed with the patent office on 2014-05-08 for system, method and data structure for fast loading, storing and access to huge data sets in real time.
This patent application is currently assigned to JETHRODATA LTD.. The applicant listed for this patent is Boaz Raufman. Invention is credited to Boaz Raufman.
Application Number | 20140129530 14/129702 |
Document ID | / |
Family ID | 47424615 |
Filed Date | 2014-05-08 |
United States Patent
Application |
20140129530 |
Kind Code |
A1 |
Raufman; Boaz |
May 8, 2014 |
SYSTEM, METHOD AND DATA STRUCTURE FOR FAST LOADING, STORING AND
ACCESS TO HUGE DATA SETS IN REAL TIME
Abstract
A computerized system including a processor and a
computer-readable non-transient memory in communication with the
processor, the memory storing instructions that when executed
manage a novel data structure and related group of algorithms that
can be used as a method for representing a set and as a base for
very efficient indexing, hash and compression. SHB is an
improvement of hierarchical bitmap. An improved database system
that can utilize the innovative data structure which includes a raw
data stream provided to the system via a data processing module,
data blocks, fields indexes tables and a keys table. There is
provided an index creating process and a columns creating process,
for transforming the data blocks and tables into index blocks and
data columns.
Inventors: |
Raufman; Boaz; (Tel-Aviv,
IL) |
|
Applicant: |
Name |
City |
State |
Country |
Type |
Raufman; Boaz |
Tel-Aviv |
|
IL |
|
|
Assignee: |
JETHRODATA LTD.
Kfar Saba
IL
|
Family ID: |
47424615 |
Appl. No.: |
14/129702 |
Filed: |
June 27, 2012 |
PCT Filed: |
June 27, 2012 |
PCT NO: |
PCT/IL2012/050222 |
371 Date: |
December 27, 2013 |
Related U.S. Patent Documents
|
|
|
|
|
|
Application
Number |
Filing Date |
Patent Number |
|
|
61457877 |
Jun 27, 2011 |
|
|
|
Current U.S.
Class: |
707/693 |
Current CPC
Class: |
G06F 7/78 20130101; G06F
16/2365 20190101; G06F 16/901 20190101; G06F 16/2272 20190101 |
Class at
Publication: |
707/693 |
International
Class: |
G06F 17/30 20060101
G06F017/30 |
Claims
1. A computerized system comprising: a processor; and a
computer-readable non-transient memory in communication with the
processor, the memory storing instructions that when executed
manage a data structure representing a set of integers that
includes: (a) at least one word, wherein each said word contains a
predefined number of, bits, wherein each said bit is selected from
the group including 1-bits and 0-bits; (b) a plurality of bit
vectors, each said bit vector containing at least one word, wherein
said at least one word is selected from the group including an
empty word containing only said 0-bits and a non-empty word
containing at least one said 1-bit; (c) a plurality of layers, each
said layer includes one said bit vector, wherein said plurality of
layers are organized sequentially; (d) the set of integers
containing a plurality of positive integer members, wherein each
member is represented by a 1-bit in a last layer and wherein the
position of each said 1-bit in said last layer is equal to a value
of said positive integer member. (e) each said non-empty word is
represented by a 1-bit in a previous layer such that a number of
said 1-bits is equivalent to a number of said non-empty words in a
subsequent said layer and a position of each of said 1-bit in said
previous layer represents a position of each said non-empty word in
a subsequent said layer; (f) a plurality of compressed layers
replacing said plurality of layers wherein said layers other than a
first said layer is compressed to a said compressed layer by
removing all of said empty words from said plurality of layers such
that each position of said removed empty words in a said layer is
represented by a position of each said 0-bit in said previous said
layer, so that said compressed layer is decompressed by calculating
said position of said removed empty words according to said
position of said 0-bits in said previous layer; (g) a said
compressed layer situation sequentially before a subsequent said
compressed layer wherein said 0-bits of said removed empty words,
of said compressed layer situated sequentially before said
subsequent compressed layer, are calculated when decompressing said
subsequent compressed layer; and (h) a plurality of counter
vectors, each of said counter vectors related to each of said
plurality of layers, wherein for each said word in each of said
layers there exists a related counter member and wherein each said
counter member holds a counter value which equals a cumulative
number of 1-bits, said cumulative number calculated from a first
position in each of said bit vectors to each respective said word
in said bit vector related to said counter member.
2. The computerized system of claim 1, wherein the data structure
is created from set of integers using a Super Hierarchiel Bitmaps
(SHB) Create function.
3. The computerized system of claim 1 wherein a positive integer
member is searched using a Super Hierarchiel Bitmaps (SHB) Search
function.
4. A machine-readable storage device having machine readable
instructions tangibly stored thereon which, when executed by the
machine, cause the machine to perform a method of database
management, the method comprising the steps of: (a) receiving, at a
data loader process, a row of values of a raw data record; (b)
decomposing, at said loader process, said row of values to separate
field values, said separation being dictated by a data scheme; (c)
assigning a numerical value key id to each said field value of said
row of values; (d) storing in a keys table one unique instance for
each said key id; (e) storing sequentially each said key id in a
pre-allocated computer readable data memory block, said data block
including rows divided into fields according to said data scheme
separation, each said row having a record id indicating a position
of said row in said data block within said plurality of data
blocks; (f) storing field indexes in a field indexes file, each
said field index being equivalent to a list of unique instances of
each said key id found in given said separate field in a given said
data block; (g) converting said field index to a Super Hierarchiel
Bitmap (SHB) format; (h) generating inverted indexes, including
said field indexes and, for each unique instance of each said key
id for each said separate field, a vector of said record ids of
said rows in which said key ids equivalent to said field index are
stored; and (i) allocating a new said data memory block when a
current row id of a said row of said key ids exceeds a preallocated
number of rows in said data block.
5. The machine-readable storage device of claim 4, wherein one
unique instance of each of said raw data values is stored in said
keys table, and wherein a said assigned key id of a said raw data
value is equal to a position of said raw data value in said keys
table.
6. The machine-readable storage device of claim 4, wherein each of
said raw data values are selected from the group of data types
including: an integer, and a non-integer.
7. The machine-readable storage device of claim 6, wherein one
unique instance of each of said non-integers is stored in said keys
table, and wherein said assigned key id of a said non-integer is
equal to a position of said non-integer field value in said keys
table.
8. The machine-readable storage device of claim 6, for a given
non-integer field value, a SHB Hash mechanism returns an existing
key id for said given non-integer field value.
9. The machine-readable storage device of claim 4, wherein said
field index further includes a counter storing a number of
instances of said key ids in each said separate field.
10. The machine-readable storage device of claim 4, wherein said
converting uses a data structure selected from the group including:
a tree, and a list, said list being ordered before said
converting.
11. The machine-readable storage device of claim 6, wherein said
key id of a said integer raw data value is derived using a
reversible function.
12. The machine-readable storage device of claim 11, wherein said
reversible function includes adding a predefined constant to said
integer raw data value.
13. The machine-readable storage device of claim 4, wherein said
raw data record is in a csv format.
11. The machine-readable storage device of claim 4, wherein said
inverted indexes are created by: (i) reading, sequentially, for
each said separate field, said rows of said key ids stored in said
data memory block; (ii) providing a vector for each unique instance
of a said key id in a said separate field in a said data block,
said vector including a list of record ids of each of said rows
which hold said key id in said separate field of a said data
block.
12. A computerized system comprising: a processor; and a
computer-readable non-transient memory in communication with the
processor, the memory storing instructions that when executed
maintain a database management system that includes: (a) a
data-processing module configured to receive a plurality of rows of
raw data values and decompose said rows of raw data values into
separate fields of raw data values, wherein said data-processing
module is further configured to assign a numerical value key id to
each said raw data value; (b) a plurality of Data Blocks each
including a plurality of said key ids sequentially stored in rows
according to said separate fields; (c) a plurality of logical
record ids, each said record id being for each of said key id rows,
said record id having a value equal a sequential position of said
row of key ids in said plurality of Data Blocks; (d) a plurality of
Data Columns, each Data Column holding a plurality of said key ids
of a given said separate field of a said Data Block; (e) a
plurality of Inverted Indexes Blocks, each said Inverted Indexes
Block having stored a plurality of ordered lists of said plurality
of record ids and each said ordered list relating to each unique
key id in a said separate field; (f) a keys table, including a list
of one unique instance of each said raw data value, wherein a
sequential position of said raw data value is equivalent to a said
assigned key id of said raw data value; and (g) a plurality of
field indexes, each said field index includes a list of unique
instances of each said key ids of a said separate field of a said
data block.
13. The computerized system of claim 12, wherein each said Data
Block is a pre allocated memory block containing a predefined
number of said rows of key ids, and each said row of key ids has a
predefined size.
14. The computerized system of claim 12, wherein each of said raw
data values are selected from the group of data types including: an
integer, and a non-integer.
15. The computerized system of claim 12, wherein said plurality of
ordered lists of said plurality of record ids are stored in a
vector.
16. The computerized system of claim 12, wherein said plurality of
ordered lists of said plurality of record ids are stored in a
Compressed Hierarchical Bitmap (CHB).
17. The computerized system of claim 12, wherein said relationship
is characterised as a mapping relationship.
Description
[0001] The present application claims priority from, and the
benefit of, U.S. Provisional Application No. 61/457,877, "System
Method and Data Structure for Fast Loading, Storing and Access to
Huge Data Sets in Real Time," filed Jun. 27, 2011. The disclosure
of the U.S. Provisional Application is hereby incorporated by
reference in its entirety.
TECHNOLOGY FIELD
[0002] The present invention deals with optimizing storage and
retrieval of data from huge databases.
BACKGROUND
[0003] Providing business intelligence over huge data volumes is an
overwhelming challenge. Organizations generate and collect
terabytes of transactions, log records, event data, traffic
information, textual messages, transcribed conversations and more
mission critical information every day, but struggle when
attempting to analyze, search and gain intelligence based on the
data. The task is becoming more and more challenging as the typical
data warehouse usage model is changing. With rapid growth in the
volumes of aggregated data, increasing number of users, new types
of queries and a growing need for real time or near real time
analytics, the legacy relational database management systems
(RDMBS) based data warehouse solutions are becoming more and more
inadequate for the changing needs in the online analytical
processing (OLAP) domain.
[0004] The scenarios of organizations dealing with billions of
events generated or collected daily are becoming more and more
common.
[0005] The legacy technology underpinning the majority of the data
warehouse market carries many deficiencies such as high time and
resources consuming load processes, high analytic latency, long
queries execution time, limited analytic capabilities, non-linear
scalability, growing hardware costs and growing maintenance
costs.
[0006] Bitmaps variants are typically used to implement inverted
indexes ids lists when there is a need to maintain and operate over
very large and usually dense ids sets. Hierarchical bitmaps and
moreover compressed hierarchical bitmaps are less commonly used,
partly due to the complexity of the implementation, and while CHB
enables to operate efficiently over spare sets as well as dense
sets it is considered wasteful for small sets. Its main advantage
is efficient union and intersection operations over large sets. In
addition to complexity of implementation hierarchical bitmaps were
also considered less optimized for disk reads and writes due to the
fact that the entire data structure must be loaded into memory for
retrieval and dump into the disk entirely for each update.
[0007] This invention brings the SHB structure that revolutionize
hierarchical bitmaps from a data structure for lists intersection
and union into full index for sets of any size (even huge ones)
that enable O(1) search time with extremely low memory requirements
compare to legacy index structure such as B-Tree, binary search
tree and hash. SHB is a most efficient indexing method that may be
used, but not limited to, as a full index for database and
generally as quick access index to elements in huge data sets.
SUMMARY
[0008] The invention may be embodied as a serial database system
designed for quick load, efficient store and quick access for huge
data sets. The system includes a data container saved in computer
readable memory/files holding data set records, possibly in encoded
format ordered by their loading order. Each record in the container
may be referred to by a logical record id which is a number
representing the sequential position of the record in the
container. Each unique value in each record column may be encoded
to a unique encoded value called key id. For each unique value or
unique key id in each column an Inverted index is created, wherein
the inverted index is an ordered list of record ids at which the
unique value is found and wherein inverted indexes are represented
as a simple vector or as CHB if the size of the index is greater
than a given threshold; and wherein inverted indexes are saved in
computer readable memory/file. The database system includes one or
more pairs of data container file and inverted indexes file;
wherein the files of each type may be merged by a background
process to larger containers holding larger range of record as long
as the above conditions are kept.
[0009] The invention may be embodied as a method of loading raw
data into a serial database system. The method includes: streaming
row data records, composed of one or more columns, to a data loader
program, wherein the row data may be in a csv format or any other
predefined agreed format; adding the record of values sequentially
to a pre-allocated computer readable data memory block, wherein the
values may be encoded to key ids and wherein said encoding may be
based on SHB hash search and insert algorithms; for each predefined
number of records to be called "transaction size", adding the
current batch of records to the end of a data container file,
wherein a single sequential write IO action is guaranteed each time
a new records batch is added to the store; for each predefined
number of records to be called "block size" (>=transaction
size), creating inverted indexes for these record called block's
inverted indexes, wherein the inverted index are created per
column. The inverted indexes block is a container for the columns
inverted indexes. The inverted indexes may be created as follows:
[0010] a. reading the records in the data memory block
sequentially; [0011] for each column/field in the record, for each
new unique value or unique key id, adding a new key to column's
keys index that may be but not limited to tree format or SHB index,
wherein in case SHB index is used a preparation process is claimed;
and wherein each key in the keys index is mapped to a list of
record ids (i.e. inverted index); and adding the current record id
to the list. [0012] For each unique value of unique key id already
in the column's keys index--adding the current record id to the end
of the record ids list [0013] If size of record ids lists is
greater than a given threshold the list may be converted to CHB;
this can be checked while populating the inverted indexes or after
all records in the block were processed [0014] b. saving the
block's inverted indexes to a file; and [0015] c. save the columns'
keys indexes to an index file that contains the column key indexes
with a mapping of their related inverted index position in
file.
[0016] The invention may also be embodied as a method of optimizing
the creation of an inverted index block by pre-built index. (This
method requires that in the load process as described above for
each record, for each unique value in the record column(s) a unique
encoded value (key id) is generated.)
[0017] For each record processed in the load process, for each
column in a record, the method includes: [0018] inserting the value
or the encoded value (key id) to a tree index; [0019] mapping each
key id to occurrences counter that increases by one every time a
key id is repeated; [0020] if the number of unique key id per
column is greater than a predefined threshold, using a key ids list
instead of a tree index; [0021] for each predefined number of
records to be called block size (>=transaction size), converting
the tree indexes and the lists if any to SHB indexes, these indexes
are called field indexes; [0022] allocating memory for column
inverted index memory block according to total number of
occurrences calculated previously, using exact allocation if exact
number of occurrences for each key is known--i.e. tree index was
used to build field index; otherwise, if a list was used,
allocating by estimated average inverted index size.
[0023] The method further includes: if exact allocation was used
when populating pre-allocated inverted indexes vectors, calculating
memory location of the inverted indexes vectors using the fields
SHB indexes;
if estimated allocation was used when populating pre-allocated
inverted indexes vectors, calculating memory location of inverted
indexes vector using the serial number of the key in the SHB index
multiplied by estimated average inverted index vector size; and if
there is a vector overflow, using temporary tree index for the
overflowed vectors. [0024] The invention may be embodied as a
method of handling update instructions in a serial database. [0025]
The method includes: [0026] fetching the row id(s) of the rows to
be updated, by executing a query based on the update instruction
filter condition; [0027] receiving row ids set from the query as
CHB; [0028] allocating new row ids to the rows containing the
updated values as if they were new inserted rows; [0029] when the
transaction containing update instructions is committed, locking
and fetching the deletion map, wherein the deletion map is a CHB
format set or rows ids that were deleted from the database; and
wherein locking is used to prevent queries execution during the
update of the deletion map; [0030] saving the rows with the updated
values in the normal process as part of the data block containing
inserted rows; [0031] merging the original rows ids set of the
updated rows (in CHB format) with the deletion map set (in CHB
format) by executing OR operation over CHB; replacing the previous
deletion map with the new deletion map; acknowledging commit of the
records batch; and unlock the deletion map. [0032] The method may
be embodied as a method of handling delete instructions in a serial
database. The method includes: [0033] fetching the row id(s) of the
rows to be deleted, by executing a query based on the delete
instruction filter condition; [0034] receiving row ids set from the
query as CHB; [0035] when the transaction containing delete
instruction is committed, locking and fetching the deletion map,
wherein the deletion map is a CHB format set of rows ids that were
deleted from the database, and wherein locking is used to prevent
queries execution during the update of the deletion map; [0036]
merging the newly deleted rows ids set (in CHB format) with the
deletion map set (in CHB format) by executing OR operation over
CHB; [0037] replacing the previous deletion map with the new
deletion map; confirming commit of delete operation; and unlocking
the deletion map.
[0038] The invention may be embodied as a method of querying and
fetching data from a serial database where the query condition is
that column C is equal to a value V. The method includes: for each
inverted index block: (a) finding key V in column C inverted index
keys and getting mapped value which is the position on the inverted
index record in the inverted index block; (b) fetching the inverted
index record from inverted index block; and (c) if inverted index
record is not in CHB structure converting in memory to CHB format;
merging inverted index CHB sets from all blocks to a single CHB
result set; in case deletion map exists, fetching deletion map and
executing the Boolean operation NOT between the result CHB and the
deletion map CHB; the result CHB is a set containing serial numbers
of the records (row ids); encoded records may be fetched from data
blocks serial storage or from alternative storage by position,
wherein records position is calculated according to row id; records
may be decoded to values using key id reversed function or key ids
index; and result of some aggregation functions such as counting
the number of records fulfilling the condition can be calculated
from the result CHB structure.
[0039] The invention may be embodied as a method for querying and
fetching data from a serial database, where the query condition is
a complex expression with two or more operands using AND, OR and
NOT operators. (A search for complex expression with two or more
operands using AND, OR and NOT operators is done by executing the
relevant Boolean operation over two or more inverted indexes in CHB
format.) The method includes: fetching inverted indexes of each
operand as described above in method for search phases A and B;
placing the inverted indexes in a stack; fetching inverted indexes
from the stack and executing the condition Boolean operations
between the inverted indexes in CHB according to the expression
condition; in case deletion map exists, fetching deletion map and
executing the Boolean operation NOT between the result CHB and the
deletion map CHB; the resulting CHB is read as a list of record ids
containing serial number of records (row ids); fetching records
from storage by position if required; and result of some
aggregation functions such as counting the number of records
fulfilling the condition can be calculated from the result CHB
structure.
[0040] The invention may be embodied as a method querying and
fetching data from a serial database where the query condition is
that column C is either greater than, greater equal than, smaller
than or smaller equal than a value V. The method includes: getting
key ids of the value range either by using values to keys reversed
function or by fetching from key ids table. Wherein a Key ids tree
index where keys are the values may be maintained for optimization
of range queries; each key in the range is considered as operand in
a complex query where all query operators are OR operators; and
continuing as with query over complex expression where all range
keys are considered as query operands for OR expression.
[0041] The invention may be embodied as a method for sorting
queries results in a serial database. (This method enable sorting
queries result according to a certain column(s) values order and
presenting the rows from the results in sorted order.) The method
includes: executing the query based on the query where condition;
fetching inverted index of the first value by values order of the
first sort by column, wherein the values order can be calculated
either from sorted key ids index or by key ids order in case keys
ids order is equal to values order (for example: if key id equal to
numeric values in numeric column); if value of inverted index is
not in CHB format, converting it to CHB format; executing
intersection operation (AND) between query result CHB and sort by
value fetched CHB, wherein the resulting CHB contains the list of
rows id to be first in the sort order, wherein rows are fetched
either from data blocks or alternative store method and preset
according to requested columns list; and wherein in case number of
rows presented is less than the number of requested rows, getting
the inverted index of the next value and repeating this step; and
if more than one sort by columns exists--repeat this using result
of the intersection as query result and fetch first/next inverted
index of the next sort by column. This is performed
recursively.
[0042] The invention may be embodied as a method for performing
aggregation queries with serial data base. (Aggregation queries are
summaries based on values of one or more columns. Aggregations are
calculated by serially scanning the original values of the
aggregated columns. In a serial database it is usually most
efficient to use the compressed column store to generate in memory
vectors of original values that match the query conditions.) The
method includes: executing the query based on the query where
condition; fetching compressed key ids from compressed column(s)
that participate in the aggregation according to row ids from
result CHB; decompressing the key ids from column; converting key
ids to original values using keys reverse functions or keys index;
placing the original values in vectors; and calculating the
aggregation over the vectors of the original values.
[0043] The invention may be embodied as a data structure having: a
sequence of bit vectors arranged in layers called levels; each K
bits in each bit vector are called a word; each bit in each vector
represents a word (K bits) in the consecutive level bit vector.
words that contain only zero bits are called empty word. Empty
words are not stored in the bit vectors with the exception of the
top level bit vector that may contain empty words; given M.sub.0
the number of words allowed in the top level bits vector and L the
number of levels the maximum numeric value in the set is M.sub.0*K
L; a counter vectors per each bit vector level wherein each counter
vector member holds a number that represents the total cumulative
number of bits counted in each position of the bit vectors starting
from the first position of the bit vector. Given that B.sub.lx be
the number of bits in a word at position x at the bit vector of
level l and C.sub.lx be the value at position x at the counter
vector of level l then C.sub.lx=C.sub.lx-1+B.sub.lx; and optional
optimization of counter vector memory sized can be achieved by the
following optimization method: multiple counter vector layers are
used for each bit vector, where each layer uses the next layer set
as a base to calculate the cumulative counter.
[0044] The invention may be embodied as a method of creating a SHB
index. The method includes: inserting index pairs (set members,
mapped values) to a tree or sorted list format, whereby the set
becomes an ordered set; converting the ordered set to SHB index
using SHB Create; allocating a new vector for mapped values. Size
of vector=size of set; and for each pair (member, mapped value) in
the ordered set fetched by order: [0045] pair.rarw.get next pair
from ordered set by order [0046] ValuesVector[n]=pair.value [0047]
n=n+1
[0048] The invention may be embodied as a method of creating a SHB
index. The method includes: inserting index keys (set members) to
an uncompressed bits vector format; converting the ordered set to
SHB index using SHB Create From BitVector; allocating a new vector
for mapped values. Size of vector=size of set; and for each pair
(member, mapped value) in the original unordered set: [0049]
SerialNo.rarw.GetSerialNumber(SHB, pair.member) [0050] Values
Vector[SerialNo-1]=pair.value
[0051] The invention may be embodied as a method of Adding new
members to a SHB index, comprising: collecting new pairs in
unordered vector set format; upon reaching a predefined threshold
converting unordered vector to SHB index+values vector; merging new
SHB index with existing SHB index (operator OR); and merging
existing and new mapped values vectors to vectors ordered by the
merged SHB index members order.
[0052] The invention may be embodied as a method of searching a SHB
index, wherein to find value V for set member M in SHB index:
[0053] SerialNo.rarw.GetSerialNumber(SHB, M) [0054] V.rarw.Values
Vector[SerialNo-1]
[0055] The invention may be embodied as a method of creating hash
using SHB index. The method includes (assuming keys collection K of
size M.): hashing keys in the set with hash function f to numeric
value within the predefined large range. If key hash value already
exists rehash until unused hash value is found; collecting the keys
and their hashed values in vectors so that the position of key k in
keys vector K is equal to the position of its hashed value in the
hashed values vector H, i.e. H[m]=.theta..sup.n (K[m]); converting
hashed values vector to SHB; and creating positions vector V as
follows: for each member H[m] in H let s be the serial number of
H[m] in the SHB set (s=getSerialNo(SHB, H[m])) and determine
V[s]=m. This means that if k is in K there exists n that denotes
number of hashes based on .theta. starting from 1 so that
K[V[getSerialNo(SHB,f.sup.n(k))]]=k, wherein the formed structure
is now a SHB index in which hashed value are the index keys and the
positions of the hashed keys in the keys vector are the mapped
values.
[0056] The invention may be embodied as a method of using hash
using SHB index. The method includes: to check if a key k is in the
hash: hashing k to a hash value; searching hash value using SHB
index search; if the hash value is found in the index, then
comparing the key to the mapped value, and if the mapped value is
not equal to the key, then rehashing and repeating search until
match is found or until the hash key is not found in the index; if
the hash key is not found in the hash index it may be added to the
hash by adding the new key to keys array K, adding it's hash value
to the SHB index and inserting its position in K to the appropriate
position in V.
[0057] The invention may be embodied as a method of SHB data
compression. The method includes: converting an array of numeric
values to be compressed to SHB set representation, wherein this SHB
set may be called the SHB compression map; determining the number
of bits representing each value in the compressed array
(bitsSize.rarw.Log 2(size(SHB))); generating the compressed array
as follows: [0058] For each numeric value in the original array
[0059] compressedValue.rarw.GetSerialNumber(SHB, Value)-1 [0060]
concatenate(compressedBitsVector, compressed Value); and preparing
a reverse index for decompression as follows: [0061] For each
member in SHB set (read in sequential order) [0062]
ReverseIndex[counter++]=member.
[0063] The invention may be embodied as a method of SHB data
decompression. The method includes: fetching the original value of
an element by its position in the original vector by: [0064]
compressed Value.rarw.getbits(compressedBitsVector,
position*bitSize, bitSize) [0065] original
Value.rarw.ReverseIndex[compressedValue] [0066] getbits(bitsVector,
pos,len).fwdarw.return len bits from bits Vector starting from
position pos
[0067] The invention may be embodied as a method for adaptive
optimization of a serial database for queries including two or more
operators. The method includes: searching query expression and
parts of the expression in a saved queries results table; if
expression found as a saved results key, fetching results from the
table; and if saved result are found it might not include row ids
that were loaded after the result was save in the table. Therefore,
if the last row id in the result fetched from table is smaller than
last row id committed, the method includes: fetching inverted
indexes of the expression operators from the inverted indexes
blocks. It is necessary to get only the part of the inverted index
that contains row ids that are not in the saved result. This can be
achieved by fetching only from inverted indexes blocks at which the
last row ids is greater than saved result last row id; executing
the Boolean operation over the fetched inverted indexes; gluing or
merging results from the saved results with the updated results
that include the newer row ids; updating saved results table with
the merged result; and using merged result as expression result.
The query execution continues, each expression in the saved results
are searched, and if not found fetch it in the standard method via
the inverted indexes. There are cases where expressions and results
might be saved in saved results table: when query execution
completed and query expression is likely to be repeated in future
queries, for example if expression contains small number of common
operators; when query execution completed and part of the query
expression is likely to be repeated in future queries, for example
if expression part contains small number of common operators.
saving only the relevant part; if query or expression are expected
to be commonly used; and if expression contains low cardinality
operands.
[0068] The invention may be embodied as a method of reordering a
serial data storage in a column data storage to enable more
efficient data access and data fetch in some queries. The method
includes: waiting for the current data block to be fully committed;
compressing each value of each column by SHB compression method
using the per column fields SHB as a compression map, wherein the
compressed value of each key id is the serial number of this key id
in the fields SHB; saving compressed bits vector in memory; for
each predefined number of key ids added to the compressed bits
vector, saving the vector a file; after all the rows in a data
block are processed into compressed column store vectors, storing
the column store vectors to disk; and deleting the current data
block from memory if it is not used by another procedure running in
parallel such as the index creation procedure.
BRIEF DESCRIPTION OF THE DRAWINGS
[0069] Various embodiments are herein described, by way of example
only, with reference to the accompanying drawings, wherein:
[0070] FIG. 1 is a prior art data structure;
[0071] FIG. 2 is a prior art data structure;
[0072] FIG. 3 is a prior art data structure;
[0073] FIG. 3A is a preferred embodiment on the inventive data
structure;
[0074] FIG. 4 is a a schematic block diagram showing the system
100;
[0075] FIG. 5 is a flowchart detailing steps for creating a
database according to the present invention;
[0076] FIG. 6 is a flowchart detailing steps for creating a
database according to the present invention;
[0077] FIG. 7 is a flowchart detailing steps for creating a
database according to the present invention;
[0078] FIG. 8 is a flowchart detailing steps for creating a
database according to the present invention;
[0079] FIG. 9 is a block diagram of an exemplary computer system
upon which the current innovation can be embodied.
DETAILED DESCRIPTION
[0080] The present invention aims to overcome the deficiencies of
prevailing data warehouse technologies by providing a revolutionary
DBMS (Database Management System) technology that enables extremely
rapid real time search and analytics over enormous volumes of
aggregated data streaming in high rate. The new DBMS technology
provides is a complete, effective and efficient solution to the big
data challenge. It is an enabling solution that offers new
capabilities through performance and capacity.
[0081] The new DBMS technology outperforms leading industry
management tools with order of magnitude query performance, faster
load rates, near zero analysis latency, fraction of hardware cost
and linear scalability. Real time access to huge data volumes which
is not a feasible option with legacy database and data warehouse
tools can be made available with the new DBMS technology.
[0082] The new DBMS technology was designed to handle a specific
task of providing real time intelligence over high streams of
events. It provides an excellent and highly efficient platform for
analytic data warehousing and is designed with a focus on the next
generation data analytics needs.
[0083] The invention includes a data base management system with
methods for loading, storing and accessing huge data sets
efficiently. The system introduces a new data structure and a new
type of index that enables fixed O(1) access time to elements in
huge data sets while utilizing economic memory space requirements
of .about.O(n) where n is the size of the set, and a new
compression method based on the invented data structure and
index.
[0084] The inventive database management system is called a "serial
database". The inventive data structure with its related algorithms
is called "super hierarchical bitmaps (SHB)".
Super Hierarchical Bitmaps (SHB)
Standard Hierarchical Bitmaps (HB)
[0085] This explains the concept of hierarchical bitmaps. SHB is an
improvement of the standard hierarchical bitmaps (HB).
[0086] A HB can be viewed as a sequence of bit vectors arranged in
layers called levels. Each bit in each vector represents a range of
k bits in the consecutive level bit vector. Let l.sub.n and l.sub.m
denote the serial numbers of two levels starting from 0 for the top
level and given that l.sub.m>l.sub.n. Let v.sub.n and v.sub.m
denote the bit vectors for levels l.sub.n and l.sub.m respectively.
Let s.sub.n and s.sub.m denote the size (number of bits) in v.sub.n
and v.sub.m respectively. Each bit in vector v.sub.n represents a
range of k (l.sub.m-l.sub.n) bits in v.sub.m and the size s.sub.m
is equal to s.sub.n*k (l.sub.m-l.sub.n).
[0087] FIG. 1 shows a hierarchical bitmaps example for L=3 and
k=4.
Compressed Hierarchical Bitmaps (CHB)
[0088] Sparse sets will result in highly space consuming HB.
Therefore a compression method is used that make HB an efficient
method to maintain bits vectors in terms of space requirements. In
a compressed Hierarchical Bitmap (CHB) we call a group of K bits a
word. We call words that contain only zero bits empty word. Empty
words are not stored in the bit vectors, resulting in much shorter
bit vectors for sparse sets. The original position of each word is
calculated according to the position of the bit representing this
word in the previous level. The original position of each bit is
calculated based on the position of the word in which the bit is
set.
[0089] The set maximum value is determined by the number of levels
in the CHB. Given number of levels L the maximum numeric value in
the set is K L. An index based on CHB is commonly a hybrid of CHB
and another index structure such as a tree or a list to manage the
CHB root nodes.
[0090] FIG. 2 shows a compressed hierarchical bitmaps example for
L=3 and k=4.
Usage of Compressed Hierarchical Bitmaps (CHB):
[0091] Adding a new element to the CHB representation of a set is
efficient, providing two conditions are met: (1) the new element
represents a numeric value greater than the numeric value of the
highest value already in the set; and (2) the memory required for
adding at most one additional word for each level bit vector is
already allocated.
[0092] CHB is an efficient method for maintaining bit vectors that
represent sparse sets. Space requirements for a set of n elements
is less than 2*(n/k) words. CHB enables calculating sets union,
sets intersection and sets difference over two or more CHB
operators in order of .about.O(n), given a small L (<100) and a
big enough K (>4). It is done by execution of the set operation
first over the top level word and then continues to the next level
using only the words from each CHB operator that are represented by
the result of the previous level operation and continue recursively
for all words in all levels.
Limitations of CHB:
[0093] CHB does not provide an efficient method for search of a
random element in the set and for fetching the serial number of a
random element in the set. Search for random element is equal to
intersection of CHB with a CHB that represents a set with the
single search element (order of .about.O(n)). Getting the serial
number of a random element is done by first searching the set and
then counting the bits in the last level preceding the searched
bit.
[0094] The inefficiency of search operation generates inefficiency
for adding a new element in the "middle" of the set (a new member
value is lower than the set highest member value). Such operation
requires a two-steps operation: first, searching for the expected
location of the new member in the CHB, and second, if new words
need to be inserted to the bit vectors then reallocation and memory
coping are needed. Minimum order due to search order:
.about.O(n).
Super Hierarchical Bitmaps (SHB)
[0095] The super hierarchical bitmap is a novel data structure and
related group of algorithms that can be used as a method for
representing a set and as a base for very efficient indexing, hash
and compression. SHB is an improvement of hierarchical bitmap.
[0096] The purpose of the SHB as part of the invention is to create
a data structure that allows the following: [0097] 1. Fast
.about.O(n) creation from ordered set. [0098] 2. Fast .about.O(n)
sets manipulation (union, intersection and set difference). [0099]
3. Fast .about.O(1) search of random element in a set. [0100] 4.
Fast .about.O(1) fetch of element sequential number in a set.
[0101] 5. Economic memory space requirements: less than
.about.O(2(n+2n/K))<O(3n) for K>4. Allows for maintaining
data structure in a contiguous memory space. [0102] 6. Economic
disk space requirements to reduced I/O operation for writing and
reading the sets representation from disk while enabling quick
conversion from the compressed disk format to the full in memory
format.
Notes:
[0103] 1. L is considered as a small value (<100) thus not
effecting order: O(nL).about.=O(n), O(L).about.=O(1). 2. Max memory
space is: O(2(n+n/K+n/K 2+ . . . +n/K
(L-1)))<O(2*(n+2n/K)<O(3n) for K>4
[0104] These attributes of the invention enable, but are not
limited to, a new and very efficient index type, a new and memory
efficient hashing method and a new compression method.
SHB Structure
[0105] FIGS. 3 and 3A show a super hierarchical bitmaps example for
L=3 and K=4.
[0106] SHB data structure contains a bits vectors array and a
counters vectors array. Unlike standard CHB the top level bit
vector of SHB may contain more than one word and may contain empty
words. Let M.sub.0 denote the number of words allowed in the top
level bits vector. The maximum numeric value of a SHB set member is
calculated as: M.sub.0*K L. In some case, for a very sparse map, a
linked list may be used to skip large gaps of empty words in the
top level.
[0107] FIG. 3A shows a super hierarchical bitmaps example for L=3
and K=4 including counter vectors.
[0108] For each bits vector a counter vector is allocated
accordingly. Counter vector members are integers. Counter vector
size is equal to the number of words in each bit vector. Counter
member max value must be greater or equal to maximum set size,
which is equal to maximum set member value (SHB sets are composed
of positive integer numbers). Each counter vector member holds a
number that represents the total cumulative number of bits counted
in each position of the SHB bit vectors starting from the first
position of the bit vector. Let B.sub.lx be the number of bits in a
word at position x at the bit vector of level l. Let C.sub.lx be
the value at position x at the counter vector of level l. We
calculate C.sub.lx=C.sub.lx-1+B.sub.lx. Counter vector memory
requirement reach O(C*(2n/K)) where C is the integer size where the
maximum integer value is greater or equal to maximum possible
number of bits in the bit vector Since SHB set may be used to
represent very large sets the size of a counter must be big enough
to represent very large number of bits, thus usually 32 bit or 64
bit integers are used.
[0109] In the example of FIG. 3A, the counter vectors are:
Level 0: [0],[2],[2],[3]
Level 1: [0],[2],[5],[7]
Level 2: [0],[2],[5],[9],[11],[14],[17],[18]
Calculation Example 1
[0110] Count total bits until position=7 at level=2 (not including
last word):
Formula: L[level][position]=L[2][7]=17
[0111] Memory size of a counter vector can be reduced by the
following optimization method: multiple counter vector layers are
used for each bit vector, where each layer uses the next layer set
as a base to calculate the cumulative counter.
[0112] With non-optimized counter vectors each member may contain a
numeric value which can be up to the maximum number of bits in the
bit vector. Therefore 4 byte or 8 bytes members are used. To reduce
counters memory size requirement we allocate a counter vector with
small size members, for example: 1 byte. We call this vector layer
1 counters. With smaller counters we can count the total number of
bits within a small range of bit vector words. A 1 byte counter may
count total of 256 bits which is equal to counting 256/K (256
divided by K number of bits) bit words. Denote W as the number of
bit words that may be counted in such a vector. When the W.sup.th
word is reached and needs to be counted we reset the counter to 0.
We add another layer of bigger counters, for example 2 bytes
counters. We call this layer 2. Layer 2 counts that total bits in
each group of W words. The number of members in layer 2 vector is
smaller than the number of member in layer 1 vector because layer 2
only counts total bits for the entire groups of W words and not for
each word in the group. We can continue this with layer 3, with,
for example 4 byte words, which would count only the total bit in
groups of W 2 words, etc. To calculate the total bits for a word in
position X of the bit vectors we sum the total counters from layer
1 at position X+layer 2 at position X/W+layer 3 at position X/W 2
etc. . . .
[0113] In the example of FIG. 3, the optimized counter vectors
are:
Level 0, Layer 1 (member size 4 bits): [0],[2],[2],[3] Level 0,
Layer 2 (member size 8 bits): [0],[7] Level 1, Layer 1 (member size
4 bits): [0],[2],[5],[7] Level 1, Layer 2 (member size 8 bits):
[0],[14] Level 2, Layer 1 (member size 4 bits):
[0],[2],[5],[9],[0],[3],[6],[7] Level 2, Layer 2 (member size 8
bits): [0],[11],[18]
Calculation Example 2
[0114] Count total bits until position=7 at level=2 (not including
last word):
Formula: L[level].Layer2[position/4]+L[level].Layer1[position]=
L[2].Layer2[7/4]+L[2].Layer1[7]=
L[2].Layer2[1]+L[2].Layer1[7]=11+6=17
Operations Over SHB
[0115] Create SHB from Ordered Set=SHB Create Function
[0116] In this operation ordered set of numbers is converted to SHB
data structure. We need to calculate the positions of the bits
representations of the set members in each SHB level and then to
calculate the SHB counters. The following procedure requires that
set members are added to SHB by order (smallest to largest).
[0117] Let U be the set member, let K be SHB word size and L be
number of SHB levels. We denote W.sub.l to be the position of the
bit vector word in bit vector V.sub.l at which the bit is to be set
at level l and B.sub.l to be the position of the bit within the
word V.sub.l[W.sub.l]. We denote R.sub.l to be the virtual position
of U in level l. The virtual position is the position of the bit
vector word at which the bit would have been set if the level bit
vector would have been a standard uncompressed bit vector. Let
T.sub.l be a counter of the total number of set bits in level
l.
[0118] Since level 0 is in fact a standard uncompressed bit vector
we start by setting W.sub.0 as follows:
W.sub.0=(U-1)/K L
[0119] Following this, for each level 1 we calculate:
R.sub.l+1=(U-1)/K (L-(l+1))
B.sub.l=R.sub.l+1%K
V.sub.l[W.sub.l]=V.sub.l[W.sub.l]|2 B.sub.l (the operation |
represents logical OR operation. Executing OR with 2 B.sub.l sets
the bit B.sub.l in bit vector word V.sub.l[W.sub.l] to 1) if
V.sub.l[W.sub.l] changed then T.sub.l=T.sub.l+1 (we increate the
counter only if number of bits in the word changed--that is a bit
that was unset before is now set to 1) W.sub.l+1=T.sub.l-1 (for
level 1 and up the position of the word depends on how many bits
were set in previous upper level, as each set bit in level l-1
represents a none empty word in level l)
[0120] Once all set members were added to SHB we calculate the SHB
counters. Let C.sub.l be the bits counter vector for level l. For
each level l we fetch all words in the V.sub.l bit vector
sequentially. Let n be the word position counter in bits vector
V.sub.l, we set:
C.sub.l[0]=0
And, for each word number n in vector V.sub.l:
C.sub.l[n+1]=C.sub.l[n]+countBits(V.sub.l[n]) (countBits be a
function that get bit vector word and return the number of set bits
in the word).
[0121] SHB counters may be saved in memory optimized layers vectors
as explain above.
Example
[0122] Populating an empty SHB where L=3 and K=4 with new member 5
and 10.
[0123] (Note: in this example the east significant bit is on left
end of the bit vector word)
[0124] Adding a new member U=5
TABLE-US-00001 L R.sub.l+1 W.sub.l B.sub.l V.sub.l[W.sub.l] T.sub.l
V.sub.l[ ] 0 0 0 0 1000 1 [1000] 1 1 0 1 0100 1 [0100] 2 4 0 0 1000
1 [1000]
[0125] Adding a new member U=10
TABLE-US-00002 L R.sub.l+1 W.sub.l B.sub.l V.sub.l[W.sub.l] T.sub.l
V.sub.l[ ] 0 0 0 0 1000 1 [1000] 1 2 0 2 0110 2 [0110] 2 9 1 1 0100
1 [1000] [0100]
[0126] Counting the bits:
TABLE-US-00003 L C.sub.l[ ] 0 [0] [1] 1 [0] [2] 2 [0] [1] [1]
Create SHB from Simple Bits Vector
[0127] In this case the original set is represented by a bit
vector. Concept is similar to creating from ordered set but some
shortcuts may apply: the last SHB level is not calculated because
the none empty words in the simple bits vector are identical to the
words of the last SHB level.
[0128] The procedure can be described as follows:
[0129] For each word Win simple input bit vector B, [0130] If W is
not empty [0131] Set its representing bit in level 0 according to
bit position in bit vector B [0132] Set its representing bits in
all levels except 0 and last level calculated based on non empty
words counter [0133] Set: last level next word=W
Search SHB=SHB Search Function
[0134] We calculate the expected positions of the bits representing
the searched value in each level of the SHB. If all expected bits
are set the search options returns true.
[0135] Let U be the searched value. Let K be SHB word size and L be
number of SHB levels. We denote W.sub.l to be the expected position
of the bit vector word in bit vector V.sub.l at which the level 1
bit for U is expected to be set and B.sub.l to be the position of
that bit within the word V.sub.l[W.sub.l]. We denote R.sub.l to be
the virtual position of U in level l. The virtual position is the
position of the bit vector word at which the expected bit would
have been set if the level bit vector would have been a standard
uncompressed bit vector.
[0136] Since level 0 is in fact a standard uncompressed bit vector
we start by setting W.sub.0 as follows:
W.sub.0=(U-1)/K L
[0137] Following this, for each level 1 we calculate:
R.sub.l+1=(U-1)/K (L-(l+1))
B.sub.l=R.sub.l+1% K
If not (V.sub.l[W.sub.l] & (2 B.sub.l)) then return false (the
operation & represents logical AND operation. Executing AND
with 2 B.sub.l check the bit B.sub.l in bit vector word
V.sub.l[W.sub.l]. If AND operation result is false then U is not a
member of the SHB set)
W.sub.l+1=C.sub.l[W.sub.l]+countLeftBits(V.sub.l[W.sub.l], B.sub.l)
(CountLeftBit--count how many bits are set to the left of bit
number B.sub.l) If l=last level then return true--if we reached the
last level it means that all expected bits are set, therefore U is
a member of the SHB set.
Get Serial Number of SHB Member
GetSerialNumber(SHB, Member)
[0138] Procedure is the same as search with one difference: if all
expected bits were set and last level reached, instead of returning
true to indicate that U is in set do:
return C.sub.l[W.sub.l]+countLeftBits(V.sub.l[W.sub.l],
B.sub.l)+1
[0139] This is the serial number of the set member U is the
SHB.
Examples
[0140] Searching and getting serial number from SHB where L=3 and
K=4.
[0141] (Note: in this example the least significant bit is on left
end of the bit vector word)
[0142] The searched SHB members are [5], [10]. The SHB:
TABLE-US-00004 L V.sub.l[ ] C.sub.l[ ] 0 [1000] [0] [1] 1 [0100]
[0] [2] 2 [1000] [0100] [0] [1] [2] [3]
[0143] Search if SHB has a member U=6 (false result is
excepted)
TABLE-US-00005 L R.sub.l+1 W.sub.l B.sub.l V.sub.l[W.sub.l] &
2{circumflex over ( )}B.sub.l Return 0 0 0 0 1000 & 1000 = true
No 1 1 0 1 0110 & 0100 = true No 2 5 0 1 1000 & 0100 =
false False
[0144] Get serial number of SHB member U=10:
TABLE-US-00006 L R.sub.l+1 W.sub.l B.sub.l V.sub.l[W.sub.l] &
2{circumflex over ( )}B.sub.l Return 0 0 0 0 1000 & 1000 = true
No 1 2 0 2 0110 & 0010 = true No 2 9 1 1 0100 & 0100 = true
2
Additional Operations Over SHB:
GetFirstMember, GetNextMember, GetLastMember, GetPreviousMember
[0145] The algorithms for the Get operations use position counters
for each word in the words bit vectors and for each bit within
current word, and increase or decrease their values in each step
according to the direction of the Get operation.
Set Intersections, Set Union, Set Difference
[0146] Algorithms are similar to intersection, union and difference
operation over CHB with one difference: when counting of total bits
at a certain position in a words bit vector is required counter
vector are used, thus reducing the bits count operation costs.
Implementing Efficient Index Using SHB
[0147] SHB index guarantees O(1) search time regardless of index
size. The preferred methods to efficiently create SHB index from an
unordered vector or pairs (key, value) and the method to add
members to SHB index and search SHB index are described as
follows:
Creating the index--Alternative 1 [0148] 1. Index pairs (set
members, mapped values) are inserted to a tree or sorted list
format, thus--the set becomes an ordered set [0149] 2. The ordered
set is converted to SHB index using SHB Create. [0150] 3. Allocate
new vector for mapped values. Size of vector=size of set [0151] 4.
For each pair (member, mapped value) in the ordered set fetch by
order: [0152] pair.rarw.get next pair from ordered set by order
[0153] ValuesVector[n]=pair.value [0154] n=n+1
Creating the Index--Alternative 2
[0154] [0155] 1. Index keys (set members) are inserted to an
uncompressed bits vector. The set becomes an ordered set without
sort. Uncompressed bit vector must be large enough to contain the
set [0156] 2. The uncompressed bit vector set is converted to SHB
index using SHB Create From Simple BitVector [0157] 3. Allocate new
vector for mapped values. Size of vector=size of set [0158] 4. For
each pair (member, mapped value) in the original unordered set:
[0159] SerialNo.rarw.GetSerialNumber(SHB, pair.member) [0160]
Values Vector[SerialNo-1]=pair.value
Adding New Members to the Index
[0160] [0161] 1. To add additional members to the index: collect
new pairs in unordered vector set format. Upon reaching a
predefined threshold convert unordered vector to SHB index+values
vector [0162] 2. Merge new SHB index with existing SHB index
(operator OR) and merge existing and new mapped values vectors to
vectors ordered by the merged SHB index members order
Search SHB Index
[0163] To find value V for set member M in SHB index:
SerialNo.rarw.GetSerialNumber(SHB, M)
V.rarw.Values Vector[SerialNo-1]
[0164] Example of SHB index:
[0165] Original set of mapped values:
[10.fwdarw.A],[1.fwdarw.],[15.fwdarw.C],[12.fwdarw.D],[8.fwdarw.E],[11.fw-
darw.F],[4.fwdarw.G],[5.fwdarw.H]
[0166] SHB bit vectors and counters (3 levels, 4 bits word):
Level 0: 1000 [1]
Level 1: 1111 [4]
Level 2: 1001 [2] 1001[4] 0111[7] 0010[8]
[0167] Values vector: [B],[G],[H],[E],[A],[F],[D],[C].rarw.(note:
first position is 0)
[0168] Search example: find the mapped value for key 12
serialNo.rarw.getSerialNumber(SHB, 12)=7 V.rarw.Values
Vector[serialNo-1]=valuesVector[7-1]=valuesVector[6]=D
Implementing Low Size Memory Efficient HASH Using SHB
Creating the Hash
[0169] Assuming keys collection K of size M. Keys in the set are
hashed with hash function f to numeric values within the predefined
large range. If key hash value already exists rehash until unused
hash value is found. The keys and their hashed values are collected
in vectors so that the position of key k in keys vector K is equal
to the position of its hashed value in the hashed values vector H,
i.e. H[m]=f.sup.n(K[m]). Hashed values vector is converted to SHB.
Create positions vector V as follows: for each member H[m] in H let
s be the serial number of H[m] in the SHB set (s=getSerialNo(SHB,
H[m])) and determine V[s]=m. This means that if k is in K there
exists n that denotes number of hashes based on f starting from 1
so that K[V[getSerialNo(SHB,f.sup.n(k))]]=k.
[0170] The formed structure is now a SHB index in which hashed
value are the index keys and the positions of the hashed keys in
the keys vector are the mapped values.
[0171] To check if a key k is in the hash, k is hashed to a hash
value which is searched via SHB index search. If the hash value is
found in the index, then the key is compared to the mapped value,
and if the mapped value is not equal to the key, then the key is
rehashed and search is repeated until match is found or until the
hash key is not found in the index.
[0172] If the hash key is not found in the hash index it may be
added to the SHB hash by adding the new key to keys array K, adding
it's hash value to the SHB index and inserting its position in K to
the appropriate position in V. For more efficient hash update it is
suggested to avoid adding single keys to the hash and instead to
collect several such new keys to collection K' of predefined size
M' and merge the entire new collection into the hash structure.
Formalization:
[0173] Let M donate the size of keys collection K to be hashed. Let
k.sub.m be a key in collection K. Let function F(k) return a unique
numeric hash value (rehash may be used to gain uniqueness) between
1 to N where N may be a very large number. Let h.sub.m be the hash
key of key k.sub.m. Let H denote the set of hashed keys. The size
of set H is equal to the size of K keys collection.
[0174] Let S denote the SHB index created from set H and let V
denote the vector of positions of keys in collection K ordered by
hash keys sequential order so that if h.sub.m=F(k.sub.m) and
s=getSerialNo(S, h.sub.m) then K[V[s]]=k.sub.m.
Example
[0175] Keys collection K of size 8: ([k0], [k1], [k2], [k3], [k4],
[k5], [k6], [k7]) Matching hash keys set H: ([10], [12], [5],
[11],[8], [1], [15], [4]).rarw.i.e: 10=F(k0), 12=F(k1), etc. . . .
)
[0176] SHB bit vectors and counters (3 levels, 4 bits word):
Level 0: 1000 [1]
Level 1: 1111 [4]
Level 2: 1001 [2] 1001[4] 0111[7] 0010[8]
[0177] SHB members and their serial numbers: 1(1), 4(2), 5[3],
8(4), 10(5), 11(6), 12(7), 15(8) Values vector V:
[5],[7],[2],[4],[0],[3],[1],[6].rarw.(note: first position is
0)
[0178] Check if k1 in hash:
if k1=K[V[getSerialNo(SHB, F(k1))-1]].fwdarw.if
k1=K[V[getSerialNo(SHB,12)-1]].fwdarw.if k1=K[V[7-1]].fwdarw.if
k1=K[V[6]].fwdarw.if k1=K[1].fwdarw.if k1=k1.fwdarw.true
Implementing Compression Using SHB
[0179] SHB compression is for arrays of positive numeric values.
The compression process converts an array of numbers, usually kept
as 32 bit or 64 bit each, to an array of compressed values. The
original values are placed in SHB and their serial numbers in the
SHB are used as their compressed representation. Since the SHB
serial numbers are sequential the number of bits required to
represent the compressed values is equal to log 2 of the number of
unique values in the array, which is likely to be significantly
smaller than the original 32 or 64 bits. The compressed values
occupy a smaller number of bits than the original values. The
values in the compressed array maintain their relative
position.
[0180] The SHB compression enables extraction of the original value
according to the value position in the original array in O(1) time
while decompressing only the extracted value.
[0181] The compression procedure is as follows: [0182] 1. Convert
the array of numeric values to be compressed to SHB set
representation. This SHB set may be called the SHB compression map
[0183] 2. Determine the number of bits representing each value in
the compressed array. [0184] bitsSize.rarw.Log 2(size(SHB)) [0185]
Note: size of SHB set is equal to number of unique values in the
original array. [0186] 3. To generate the compressed array: [0187]
For each numeric value in the original array [0188]
compressedValue.rarw.GetSerialNumber(SHB, Value)-1 [0189]
concatenate(compressedBitsVector, compressed Value) [0190] 4.
Prepare reverse index for decompression: [0191] For each member in
SHB set (read in sequential order) [0192]
ReverseIndex[counter++]=member
[0193] To fetch the original value of an element by its position in
the original vector:
compressed Value.rarw.getbits(compressedBitsVector,
position*bitsSize, bitsSize) original
Value.rarw.ReverseIndex[compressedValue] Note: getbits(bitsVector,
pos, len).fwdarw.return len bits from bits Vector starting from
position pos
Example of SHB Compression:
[0194] Original array: ([10], [15], [12], [1], [8], [12], [10],
[11], [4], [5])
[0195] SHB bit vectors and counters (3 levels, 4 bits word):
Level 0: 1000 [1]
Level 1: 1111 [4]
Level 2: 1001 [2] 1001[4] 0111[7] 0010[8]
[0196] Total number of unique values: 8 Number of bits representing
each compressed value=log 2(8)=3 Reverse index: ([1], [4], [5],
[8], [10], [11], [12], [15]) Compressed representation: 100 111 110
000 011 110 100 101 001 010 Translation of original values to
compressed bits representation according to SHB serial number:
1.fwdarw.[000], 4.fwdarw.[001], 5.fwdarw.[010], 8.fwdarw.[011],
10.fwdarw.[100], 11.fwdarw.[101], 12.fwdarw.[110],
15.fwdarw.[111]
Serial Database
[0197] FIG. 4 is a schematic block diagram showing the system 100
comprising a raw data stream 105 provided to the system via a data
processing module 110, data blocks 120, fields indexes tables 140
and a keys table 130, as will be explained in detail below, an
index creating process 150 and a columns creating process 160, for
transforming the data blocks 120 and tables (130,140) into index
blocks 170 and data columns 180, as will be explained in detail
below.
Schema:
[0198] A schema is provided describing the data base data schema,
tables and columns, and the format and structure of the raw data to
be loaded into the data base management system.
Data Blocks:
[0199] A data block 120 is a pre allocated memory block that
contains a set of records in encoded format, where raw data values
are encoded to numbers. The code is called key id. Key id size is
fixed and predefined (usually 4 or 8 bytes per id) and therefore
the size of a record in the data block is fixed. The original raw
data records are converted to vectors of key ids and placed in the
data block in a serial order (i.e. a loading order). The number of
records in a data block is predefined and therefore the data block
size can be pre-calculated and pre-allocated in memory as one
contiguous memory block. Data blocks are dumped to the disk, as
will be explained in detail below.
Keys:
[0200] A single keys table 130 stores all the unique keys (i.e. one
unique instance of each key id) created by the data processing
module. Each unique value may be encoded to a unique numeric value
called key id. Encoding of an integer or real number is a
reversible function returning a unique integer per unique value
(for example: where the unique integer is equal to the unique value
plus a predetermined constant [the constant can be negative or
positive, e.g. if the original value is 175 and the predetermined
constant is 5, then the key id for this value would be 180]).
Encoding of a string returns a sequential number. One unique
instance of each string is stored in the keys table. The sequential
position (i.e. a specific value) of the stored string--in the
string table--is the value of the key id of that string. String to
number indexes are stored in SHB hash, thus allowing O(1) search
time. Other non-integer data types (e.g. a date data type etc.) may
be encoded with one of the above two methods. In cases where key
ids are assigned without a reversible function, a table--mapping
key ids to values--is kept in memory and stored on disk whenever a
commit is executed. SHB hash or SHB index may be used to fetch key
id by value.
Field Indexes:
[0201] Fields indexes table 140 is built for each data block 120.
The fields indexes hold the unique keys of each column in a data
block in SHB format. Fields indexes may also include counters for
number of occurrences of each key in the data block as mapped
values to SHB index. In the fields creation process key ids are
added to a per columns index tree. The index holds the unique keys
and the counter of number of occurrences of each key in the data
block as a mapped value. In case the number of unique values is
greater than a predefined threshold, the tree structure is no
longer used to build the indexes. Instead the list of key ids is
kept in a vector, no counters are kept and values may appear more
than once in the vector. When all keys ids per data blocks are
collected to the tree index or to a vector it is converted to SHB
format and the occurrences if exists is kept as SHB index mapped
values vector.
Inverted Indexes Block:
[0202] Inverted index block is composed of unique keys index per
column and an ordered list of row ids per unique key indicating the
location of the key's original record in a data block. The fields
SHB index that contains the unique keys per column is used as the
unique keys index. The ordered lists of rows ids are generated as
vectors but may be converted to a more efficient CHB format given
that vector size is bigger than a given threshold.
[0203] The inverted indexes are created by fetching data block rows
sequentially and storing the row id in vectors related to the key
ids from the row. The creation process uses a pre-allocated
inverted index memory blocks. An inverted index memory block is
composed of per column vectors. The vectors are allocated according
to the total number of occurrences calculated when generating the
field index in the data processing step, assuming that the exact
number of occurrences for each key is known--i.e. a tree index was
used to build the field index. Otherwise, if a list was used,
allocation is done by estimation based on average inverted index
size and the number of unique keys.
[0204] Populating the inverted indexes vectors is done by
calculating the relative positions in the inverted indexes vectors
according to the serial number of the key in the fields SHB
indexes.
[0205] In case estimated allocation was used, then populating
pre-allocated inverted indexes vectors is done by calculating
position in the vector using the serial number of the key in the
SHB index multiplied by estimated average inverted index vector
size. In case the estimated pre-allocated space per key is not
sufficient to hold the inverted index of a specific key, a
temporary tree index mapped to temporary vectors is used to
complete the pre-allocated vectors.
[0206] If size of row ids list is greater than a given threshold
the list may be converted to CHB. This can be checked while
populating the inverted indexes vectors or after all the records in
the block were processed.
Example of Inverted Indexes Creation in a Pre Allocated Vector
[0207] Field index set (unique key id values) and counters (number
of occurrences):
1 [5], 4 [2], 5 [1], 8 [1], 10 [3], 11 [2], 12 [3], 15 [1]
[0208] SHB bit vectors and counters (3 levels, 4 bits word):
Level 0: 1000 [1]
Level 1: 1111 [4]
Level 2: 1001 [2] 1001 [4] 0111 [7] 0010 [8]
[0209] Positions vector: [0], [5], [7], [8], [9], [12], [14],
[17].rarw.Act as SHB index values vector Size of allocated inverted
index vector: 18 (total number of occurrences) Step 1: Adding value
U1 to inverted index of key 10: serialNo.rarw.getSerialNo(SHB,
10)=5
position.rarw.positionsVector[serialNo-1]=positionsVector[5]=9
InvertedIndexVector[position]=invertedIndexVector[9].rarw.U1
#
InvertedIndexVector.fwdarw.[|0,0,0,0,0|0,0|0|0|U1,0,0|0,0|0,0,0|0|0]
[0210]
positionsVector[serialNo-1].rarw.positionsVector[serialNo-1]+1
# Positions Vector.fwdarw.[0,5,7,8,10,12,14,17]
[0211] Step 2: Adding value U2 to inverted index of key 10:
serialNo.rarw.getSerialNo(SHB, 10)=5
position.rarw.positionsVector[serialNo-1]=positionsVector[5]=10
InvertedIndexVector[position]=invertedIndexVector[10].rarw.U1
#
InvertedIndexVector.fwdarw.[|0,0,0,0,0|0,0|0|0|U1,U2,0|0,00,0,0|0|0]
[0212] positions
Vector[serialNo-1].rarw.positionsVector[serialNo-1]+1
# Positions Vector.fwdarw.[0,5,7,8,11,12,14,17]
Creating a Database:
[0213] FIGS. 5 through 8 are flowcharts showing the sequence of
steps for creating a database or adding data to an existing
database according to the present invention.
[0214] FIG. 5 shows the acquisition of raw data. In step 200 the
system waits for a batch of records to be forwarded after fetch.
Raw data 105 is provided as stream of raw records to the loader
process 110.
[0215] Raw data 105 may also include loading instructions such as
commit or delete. In step 210 each record in the batch is
decomposed to separate column/field values, based on the data
scheme. In step 220 key ids are assigned to each value in a row, as
described above. One instance of each key id is stored in the keys
table 130. In step 230 the assigned key ids are placed sequentially
in the current data block 120. A new data block is allocated when
the current record serial number, also called row id, exceeds the
last record serial number that can be stored in previously
allocated blocks. In step 240 the key ids are added to the current
field indexes table 140, as described above.
[0216] FIG. 6 shows a process that may run in parallel to the
process of FIG. 5, where the system checks whether the current
transaction is committed. The serial number of the last processed
row is monitored to check when this number is greater or equal to
next serial number to commit Determining the next serial number to
commit can be done either by checking the serial position of the
next commit instruction in the raw data stream or by setting fixed
auto commit transaction size and calculating next serial number to
commit=last serial number committed+transaction size. A commit
action is executed once the processing of all rows in a transaction
is completed, i.e. when the data processing thread completed the
processing of all the rows in the transaction and those rows are
now placed in a data block. Committing means saving the committed
rows from the data block to the hard disk. In step 300 the system
waits for the current transaction to be committed. If it is
committed, the committed rows from the data block are saved to disk
in step 310. In step 320 the system checks whether all the data in
the current data block has been committed and returns to step 300
if not. In step 330, if it was determined that the entire data
block has been committed, the field indexes 140 of the data block
120 are converted to SHB indexes, as described above, and saved to
indexes storage 170 on disk. Once writing to the disk is completed
successfully and sync is confirmed the system may send
acknowledgment that the commit is completed. In case the committed
rows range is the last range in a data block the data block is
marked as a fully committed data block in step 340.
[0217] FIG. 7 shows a process that may run in parallel to the
processes of FIGS. 5 and 6, where the system checks whether the
current data block is fully committed. In step 400 the system waits
for the current data block 120 to be fully committed. In step 410,
when it was determined that the current data block is fully
committed, memory is allocated for inverted indexes block, as
described above. In step 420 the inverted indexes are created from
the current data block, as described above. Once all inverted
indexes per current data block have been created the inverted
indexes block is saved to the disk in step 430, along with a file
positions vector that contains inverted index vectors position in
the inverted index file. The file positions vector is ordered by
key ids.
[0218] Once all the rows in a data block were processed into
inverted indexes and the inverted indexes block was saved
successfully to disk the data block may be deleted from memory if
it is not used by another procedure running in parallel such as the
alternative store creation procedure.
[0219] The position of any inverted index vectors can be later
fetched in O(1) time by using SHB getSerialNumber function over the
fields SHB that was already created in the commit step. The
getSerialNumber function returns the serial number of a key which
matched the position in the file positions vector that holds the
file position of the inverted index.
[0220] FIG. 8 shows a process that may run in parallel to the
processes of FIGS. 5, 6 and 7, where data is stored in another
storage method in addition to the data blocks rows serial storage
described above. The purpose of this process is to create
alternative data access methods that are more efficient for the
execution of certain types of queries.
[0221] The alternative storage method may include (but is not
limited to) column store. The method described herein explains the
creation of alternative store methods that may be generated from
sequential reading of the loaded data blocks, such as column
store.
[0222] The following describe in details creation of column
store:
[0223] In step 500, the system waits for the current data block to
be fully committed. In step 510, the process of creating compressed
column store from data block begins when it has been determined
that the current data block has been fully committed. The columns
are compressed by SHB compression method using the per column
fields SHB created in the step 330 as a compression map. The
compressed value of each key id is the serial number of this key id
in the fields SHB.
[0224] The rows are fetched sequentially from the data block. Each
value of each column in each row is compressed by SHB compression
and saved in compressed bits vector in memory (step 520). For each
predefined number of key ids added to the compressed bits vector
the vector is saved to a file in step 530. A check point is used to
keep the sequential number of the last value saved successfully.
Once all the rows in a data block were processed into compressed
column store vectors the column store vectors are stored to disk in
step 530 and the current data block may be deleted from memory if
it is not used by another procedure running in parallel such as the
index creation procedure.
[0225] The sequential number or row id of the last row stored in
the alternative store is always less than or equal to the row id of
the last row committed in the database. Therefore the querying
method must use the alternative store check point to determine
which values may be fetched from the alternative store files and
which values may only be fetched from the data blocks.
Deletions and Updates
[0226] In certain scenarios a transaction may include a row delete
instruction or one or more column values update for a row. The
delete or update instruction may be part of an inserted rows batch
or may be sent separately. If the scheme includes unique keys, then
in the case that a unique key of a row within rows batch is equal
to a unique key of a row already inserted in the system the row is
considered an update row. Unique key check is done by searching the
merged fields SHB index per key.
Handling Delete:
[0227] 1. Fetch the row id(s) of the rows to be deleted. This is
done by executing a query based on the delete instruction filter
condition. Row ids set is returned from the query as CHB. [0228] 2.
When the transaction containing delete instruction is committed
then lock and fetch the deletion map. The deletion map is a CHB
format set of rows ids that were deleted from the database. Lock is
used to prevent queries execution during the update of the deletion
map. [0229] 3. Merge the newly deleted rows ids set (in CHB format)
with the deletion map set (in CHB format) by executing OR operation
over CHB. [0230] 4. Replace the previous deletion map with the new
deletion map. Confirm commit of delete operation. Unlock delete
map.
Handling Update (Also, but not Limited to, as Part of Records
Batch):
[0230] [0231] 1. Fetch the row id(s) of the rows to be updated.
This is done by executing a query based on the update instruction
filter condition. Row ids set is returned from query as CHB. New
row ids are allocated to the rows containing the updated values as
if they were regularly inserted rows [0232] 2. When the transaction
containing update instructions is committed then lock and fetch the
deletion map. The deletion map is a CHB format set or rows ids that
were deleted from the database. Lock is used to prevent queries
execution during the update of the deletion map. [0233] 3. Save the
rows with the updated values in the normal process as part of the
data block containing inserted rows. [0234] 4. Merge the original
rows ids set of the updated rows (in CHB format) with the deletion
map set (in CHB format) by executing OR operation over CHB. [0235]
5. Replace the previous deletion map with the new deletion map.
Acknowledge commit of the records batch. Unlock delete map.
Querying:
[0236] A search for records where the condition is that column C is
equal to a value V is conducted by: [0237] A) For each inverted
index block: [0238] a. Find key V in column C inverted index keys.
Get mapped value which is the position on the inverted index record
in the inverted index block [0239] b. Fetch inverted index record
from inverted index block [0240] c. If inverted index record is not
in CHB structure convert in memory to CHB format [0241] B) Merge
inverted index CHB sets from all blocks to a single CHB result set
[0242] C) In case deletion map exist fetch deletion map and execute
the Boolean operation NOT between the result CHB and the deletion
map CHB [0243] D) The result CHB is a set containing serial numbers
of the records (row ids). [0244] E) If records are requested
records are fetched from serial storage or alternative storage by
position. Records position is calculated according to row id.
[0245] F) Some aggregation functions such as count of number of
records fulfilling the condition are calculated from the result CHB
structure
[0246] A search for complex expression with two or more operands
using AND, OR and NOT operators is done by executing the relevant
Boolean operation over two or more inverted indexes in CHB format:
[0247] A) Fetch inverted indexes of each operand as described above
in method for search phases A and B and place the inverted indexes
in a stack [0248] B) Fetch inverted indexes from the stack and
execute the condition Boolean operations between the inverted
indexes in CHB according to the expression condition [0249] C) In
case deletion map exists fetch deletion map and execute the Boolean
operation NOT between the result CHB and the deletion map CHB
[0250] D) The resulting CHB is read as a list of record ids
containing serial number of records (row ids) [0251] E) Records are
fetched from storage by position if required [0252] F) Some
aggregation functions such of count of number of records fulfilling
the condition are calculated from the result CHB structure
[0253] A search over a range of record where the condition is that
column C is either greater than, greater equal than, smaller than
or smaller equal than a value V is conducted by: [0254] A) Get key
ids of value range either by using values to keys reversed function
or by fetching from key ids table. Key ids tree index where keys
are the values may be maintained for optimization of range queries.
[0255] B) Each key in the range is considered as operand in a
complex query where all query operators are OR operators. Continue
as with query over complex expression where all range keys are
considered as query operands for OR expression.
Sorting Queries Results
[0256] This method enables sorting queries result according to a
certain column(s) values order and presenting the rows from the
results in sorted order. [0257] A) Execute the query based on the
query where condition. [0258] B) Fetch inverted index of the first
value by values order of the first sort by column. The values order
can be calculated either from sorted key ids index or by key ids
order in case keys ids order is equal to values order (for example:
if key id equal to numeric values in numeric column). If inverted
index is not in CHB format convert it to CHB format. [0259] C)
Execute intersection operation (AND) between query result CHB and
sort by value fetched CHB. The resulting CHB contains the list of
rows id to be first in the sort order. Fetch rows either from data
blocks or alternative store method and preset according to
requested columns list. In case number of rows presented is less
than the number of requested rows get the inverted index of the
next value and repeat this step. [0260] D) If more than one sort by
columns exists--repeat this using result of the intersection as
query result and fetch first/next inverted index of the next sort
by column. Do this recursively. Performing Aggregation Queries with
Serial Data Base
[0261] Aggregation queries are summaries based on values of one or
more columns. Aggregations are calculated by serially scanning the
original values of the aggregated columns. In a serial database it
is usually most efficient to use the compressed column store to
generate in memory vectors of original values that match the query
conditions. [0262] A) Execute the where condition [0263] B) Fetch
compressed key ids from compressed column(s) that participate in
the aggregation according to row ids from result CHB [0264] C)
Decompress the key ids from column, convert key ids to original
values using keys reverse functions or keys index and placed the
original values in vectors [0265] D) Calculate the aggregation over
the vectors of the original values
Adaptive Optimization of Serial Database
[0266] Optimization method for serial database the improved queries
performances in adaptive manner The optimization is dynamically
adapted to both application queries need and data characteristics.
The optimization method reduces query time by accessing full or
partial queries results saved in a special adaptive optimization
saved results table that contain query expression as key (operands
and operators) and results in CHB format.
[0267] Before executing the method for query and fetch data from
serial database, if the query expression includes two or more
operators: [0268] A) Search query expression and parts of the
expression in saved results table [0269] B) If expression found as
a save results key fetch results from table [0270] C) If saved
result found it might not include row ids that were loaded after
the result was save in the table. Therefore, if last row id in the
result fetched from table is smaller than last row id committed:
[0271] a. Fetch inverted indexes of the expression operators from
the inverted indexes blocks. Need to get only the part of the
inverted index that contains row ids that are not in the saved
result. This can be achieved by fetch only from inverted indexes
blocks at which last row ids is greater than save result last row
id. [0272] b. Execute the Boolean operation over the fetched
inverted index [0273] c. Glue or merge results from save results
with the updated results that include the newer row ids. [0274] d.
Update saved results table with the merged result [0275] e. Use
merged result as expression result [0276] D) Continue with query
execution. Search each expression in the saved results. If not
found fetch it in the standard method via the inverted indexes.
[0277] Cases where expressions and results might be saved in saved
results table: [0278] 1. When query execution completed and query
expression is likely to be repeated in future queries, for example
if expression contains small number of common operators [0279] 2.
When query execution completed and part of the query expression is
likely to be repeated in future queries, for example if expression
part contains small number of common operators. Save only the
relevant part [0280] 3. If query or expression are expected to be
commonly used [0281] 4. If expression contain low cardinality
operands
[0282] FIG. 9 is a block diagram of an exemplary computer system
upon which the current innovation can be embodied. Referring now to
FIG. 9, the exemplary computer system 1100 upon which the current
innovation can be embodied, includes a processing unit 1104 that
includes one or more processors. Processor unit 1104 is connected
to a communication infrastructure 1102, which may comprise, for
example, a bus or a network.
[0283] Computer system 1100 also includes a main memory 1106,
preferably random access memory (RAM), and may also include a
secondary memory 1120. Secondary memory 1120 may include, for
example, a hard disk drive 1122, a removable storage drive 1124,
and/or a memory stick. Removable storage drive 1124 may comprise a
floppy disk drive, a magnetic tape drive, an optical disk drive, a
flash memory, or the like. Removable storage drive 1124 reads from
and/or writes to a removable storage unit 1128 in a well-known
manner Removable storage unit 1128 may comprise a floppy disk,
magnetic tape, optical disk, or the like, which is read by and
written to by removable storage drive 1124. As will be appreciated
by persons skilled in the relevant art(s), removable storage unit
1128 includes a computer usable storage medium having stored
therein computer software and/or data.
[0284] In alternative implementations, secondary memory 1120 may
include other similar means for allowing computer programs or other
instructions to be loaded into computer system 1100. Such means may
include, for example, a removable storage unit 1130 and an
interface 1126. Examples of such means may include a program
cartridge and cartridge interface (such as that found in video game
devices), a removable memory chip (such as an EEPROM, EPROM, or
PROM) and associated socket, and other removable storage units 1130
and interfaces 1126 which allow software and data to be transferred
from the removable storage unit 1130 to computer system 1100.
[0285] As used herein, the terms "computer program medium" and
"computer readable medium" are used to generally refer to media
such as removable storage unit 1128, removable storage unit 1130
and a hard disk installed in hard disk drive 1122. Computer program
medium and computer readable medium can also refer to memories,
such as main memory 1106 and secondary memory 1120, which can be
semiconductor devices (e.g., DRAMS, etc.). These computer program
products are means for providing software to computer system
1100.
[0286] Computer programs (also called computer control logic,
programming logic, or logic) are stored in main memory 1106 and/or
secondary memory 1120. Such computer programs, when executed,
enable the computer system 1100 to implement features of the
present invention as discussed herein. Accordingly, such computer
programs represent controllers of the computer system 1100. Where
the invention is implemented using software, the software may be
stored in a computer program product and loaded into computer
system 1100 using removable storage drive 1124, or interface
1126.
[0287] The invention is also directed to computer program products
comprising software stored on any computer readable medium. Such
software, when executed in one or more data processing devices,
causes a data processing device(s) to operate as described herein.
Embodiments of the present invention employ any computer readable
medium, known now or in the future. Examples of computer readable
mediums include, but are not limited to, primary storage devices
(e.g., any type of random access memory) and secondary non-volatile
storage devices (e.g., hard drives, floppy disks, CD ROMS, zip
disks, tapes, magnetic storage devices, optical storage devices,
MEMs, nanotechnology-based storage device, etc.).
[0288] Note that the instructions of the software discussed above
can be provided on one computer-readable or computer-usable storage
medium, or alternatively, can be provided on multiple
computer-readable or computer-usable storage media distributed in a
large system having possibly plural nodes. Such computer-readable
or computer-usable storage medium or media is (are) considered to
be part of an article (or article of manufacture). An article or
article of manufacture can refer to any manufactured single
component or multiple components.
* * * * *